Report Open Access

Fail-Safe Execution of Deep Learning based Systems through Uncertainty Monitoring

Michael Weiss; Paolo Tonella

Citation Style Language JSON Export

  "publisher": "Zenodo", 
  "DOI": "10.5281/zenodo.5055710", 
  "author": [
      "family": "Michael Weiss"
      "family": "Paolo Tonella"
  "issued": {
    "date-parts": [
  "abstract": "<p>Modern software systems rely on Deep Neural Networks (DNN) when processing complex, unstructured inputs, such as images, videos, natural language texts or audio signals.&nbsp;Provided the intractably large size of such input spaces, the intrinsic limitations of learning algorithms&nbsp; and the ambiguity about the expected predictions for some of the inputs, not only there is no guarantee that DNN&#39;s predictions are always correct, but rather developers must safely assume a low, though not negligible, error probability.&nbsp;A fail-safe Deep Learning based System (DLS) is one equipped to handle DNN faults by means of a supervisor, capable of recognizing predictions that should not be trusted and that should activate a healing procedure bringing the DLS to a safe state.</p>\n\n<p>In this paper, we propose an approach to use DNN uncertainty estimators to implement such supervisor.&nbsp;We first discuss advantages and disadvantages of existing approaches to measure uncertainty for DNNs&nbsp;and propose novel metrics for the empirical assessment of the&nbsp; supervisor that rely on such approaches.&nbsp;We then describe our publicly available tool Uncertainty-Wizard, which allows transparent estimation of uncertainty for regular tf.keras DNNs.&nbsp;Lastly, we discuss a large-scale&nbsp; study conducted on four different subjects to empirically validate the approach,&nbsp;reporting the lessons-learned as guidance for software engineers who intend to monitor uncertainty for fail-safe execution of DLS.</p>", 
  "title": "Fail-Safe Execution of Deep Learning based Systems through Uncertainty Monitoring", 
  "type": "article", 
  "id": "5055710"
All versions This version
Views 2727
Downloads 2828
Data volume 25.3 MB25.3 MB
Unique views 2323
Unique downloads 2525


Cite as